AReaL is an open-source reinforcement learning framework that connects large language models with agent applications. It acts as an “RL bridge”: developers can take an existing LLM-based agent and improve it with reinforcement learning, without rebuilding their application around a training framework. The project has collected more than 5,800 stars on GitHub and remained under active development in 2026.
Training agents with reinforcement learning means coordinating rollout generation, reward computation and policy updates at scale, which is where most home-grown setups break down. AReaL provides a simple and flexible architecture for this, decoupling generation from training so systems can scale from a single workstation to distributed clusters. It targets both reasoning models and interactive agent workloads, and is designed to plug into the agent frameworks and model-serving stacks teams already use, rather than replacing them. The team behind it comes from large-scale RL research, and the design reflects lessons from training production reasoning systems: fault tolerance during long runs, efficient sampling so expensive GPU time is not wasted waiting for rollouts, and reproducible experiment configuration. Documentation and examples in the repository walk through training agents end to end, from environment setup to evaluating the improved policy. Sample configurations for popular open-weight models lower the barrier for a first experiment.
Free. AReaL is open source under the Apache-2.0 license; you only pay for the compute you train on.
Machine-learning engineers and researchers who post-train LLM agents with reinforcement learning and want production-grade training infrastructure instead of research scripts.